AI-centric companies accounted for more than half of all global cybersecurity deals by late 2025. The hardest question in security DD is whether the AI actually works — or whether it generates more alerts than the customer can handle.

Behavioural AI that detects novel threats without signatures. The most durable security AI is anomaly detection and behavioural analysis that can identify previously unknown attack techniques by recognising unusual patterns in user, system, or network behaviour. This capability requires high-quality training data, continuous learning pipelines, and low-latency inference at production scale. We assess whether the behavioural model has been tested against genuinely novel threat scenarios.

AI that reduces analyst workload rather than increasing it. The security tools with the highest NRR and retention are those that demonstrably reduce the time analysts spend on false positives and low-priority alerts. AI triage, automated enrichment, and guided response playbooks that compress the analyst workflow are the features that security operations teams actually value.

Identity and access intelligence as a durable category. Identity-based attacks — credential theft, privilege escalation, lateral movement — now account for the majority of significant enterprise breaches. AI-powered identity threat detection and response sits at the intersection of IAM and security operations and addresses a threat vector that perimeter-focused tools cannot see. Companies with strong identity graph analytics and behavioural baselines have a technically differentiated position.

AI that detects yesterday's threats reliably but misses today's. ML models trained on historical threat data have a known weakness against novel attack techniques — including adversarial evasion specifically designed to avoid detection. An attacker who knows the model architecture can craft inputs that evade detection. We assess whether the company conducts adversarial robustness testing as part of its quality programme.

Alert fatigue that destroys customer trust and drives churn. A security product that generates too many false positives will be turned off or ignored by analysts — which is worse than not having a security product at all, because it creates false assurance. We assess false positive rates in production environments, not in demo or test environments.

The security platform as an attack target. Security platforms have access to the most sensitive data in an enterprise — network traffic, endpoint telemetry, user behaviour, credentials. A breach of the security platform itself would be catastrophic for the vendor's customers and the vendor's business. We assess the security posture of the platform, including its attack surface, supply chain dependencies, and controls around privileged access.